Whamcloud - gitweb
Branch HEAD
[fs/lustre-release.git] / lustre / quota / quota_context.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/quota/quota_context.c
37  *
38  * Lustre Quota Context
39  *
40  * Author: Niu YaWei <niu@clusterfs.com>
41  */
42
43 #ifndef EXPORT_SYMTAB
44 # define EXPORT_SYMTAB
45 #endif
46
47 #define DEBUG_SUBSYSTEM S_LQUOTA
48
49 #include <linux/version.h>
50 #include <linux/fs.h>
51 #include <asm/unistd.h>
52 #include <linux/slab.h>
53 #include <linux/quotaops.h>
54 #include <linux/module.h>
55 #include <linux/init.h>
56
57 #include <obd_class.h>
58 #include <lustre_quota.h>
59 #include <lustre_fsfilt.h>
60 #include <class_hash.h>
61 #include <lprocfs_status.h>
62 #include "quota_internal.h"
63
64 #ifdef HAVE_QUOTA_SUPPORT
65
66 static lustre_hash_ops_t lqs_hash_ops;
67
68 unsigned long default_bunit_sz = 128 * 1024 * 1024; /* 128M bytes */
69 unsigned long default_btune_ratio = 50;             /* 50 percentage */
70 unsigned long default_iunit_sz = 5120;              /* 5120 inodes */
71 unsigned long default_itune_ratio = 50;             /* 50 percentage */
72
73 cfs_mem_cache_t *qunit_cachep = NULL;
74 struct list_head qunit_hash[NR_DQHASH];
75 spinlock_t qunit_hash_lock = SPIN_LOCK_UNLOCKED;
76
77 /* please sync qunit_state with qunit_state_names */
78 enum qunit_state {
79         /**
80          * a qunit is created
81          */
82         QUNIT_CREATED      = 0,
83         /**
84          * a qunit is added into qunit hash, that means
85          * a quota req will be sent or is flying
86          */
87         QUNIT_IN_HASH      = 1,
88         /**
89          * a qunit is removed from qunit hash, that
90          * means a quota req is handled and comes back
91          */
92         QUNIT_RM_FROM_HASH = 2,
93         /**
94          * qunit can wake up all threads waiting for it
95          */
96         QUNIT_FINISHED     = 3,
97 };
98
99 static const char *qunit_state_names[] = {
100         [QUNIT_CREATED]      = "CREATED",
101         [QUNIT_IN_HASH]      = "IN_HASH",
102         [QUNIT_RM_FROM_HASH] = "RM_FROM_HASH",
103         [QUNIT_FINISHED]     = "FINISHED",
104 };
105
106 struct lustre_qunit {
107         struct list_head lq_hash;          /** Hash list in memory */
108         atomic_t lq_refcnt;                /** Use count */
109         struct lustre_quota_ctxt *lq_ctxt; /** Quota context this applies to */
110         struct qunit_data lq_data;         /** See qunit_data */
111         unsigned int lq_opc;               /** QUOTA_DQACQ, QUOTA_DQREL */
112         cfs_waitq_t lq_waitq;              /** Threads waiting for this qunit */
113         spinlock_t lq_lock;                /** Protect the whole structure */
114         enum qunit_state lq_state;         /** Present the status of qunit */
115         int lq_rc;                         /** The rc of lq_data */
116 };
117
118 #define QUNIT_SET_STATE(qunit, state)                                   \
119 do {                                                                    \
120         spin_lock(&qunit->lq_lock);                                     \
121         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
122                     "lq_rc(%d)\n",                                      \
123                     qunit, qunit_state_names[qunit->lq_state],          \
124                     qunit_state_names[state], qunit->lq_rc);            \
125         qunit->lq_state = state;                                        \
126         spin_unlock(&qunit->lq_lock);                                   \
127 } while(0)
128
129 #define QUNIT_SET_STATE_AND_RC(qunit, state, rc)                        \
130 do {                                                                    \
131         spin_lock(&qunit->lq_lock);                                     \
132         qunit->lq_rc = rc;                                              \
133         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
134                     "lq_rc(%d)\n",                                      \
135                     qunit, qunit_state_names[qunit->lq_state],          \
136                     qunit_state_names[state], qunit->lq_rc);            \
137         qunit->lq_state = state;                                        \
138         spin_unlock(&qunit->lq_lock);                                   \
139 } while(0)
140
141
142 int should_translate_quota (struct obd_import *imp)
143 {
144         ENTRY;
145
146         LASSERT(imp);
147         if (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64)
148                 RETURN(0);
149         else
150                 RETURN(1);
151 }
152
153 void qunit_cache_cleanup(void)
154 {
155         int i;
156         ENTRY;
157
158         spin_lock(&qunit_hash_lock);
159         for (i = 0; i < NR_DQHASH; i++)
160                 LASSERT(list_empty(qunit_hash + i));
161         spin_unlock(&qunit_hash_lock);
162
163         if (qunit_cachep) {
164                 int rc;
165                 rc = cfs_mem_cache_destroy(qunit_cachep);
166                 LASSERTF(rc == 0, "couldn't destroy qunit_cache slab\n");
167                 qunit_cachep = NULL;
168         }
169         EXIT;
170 }
171
172 int qunit_cache_init(void)
173 {
174         int i;
175         ENTRY;
176
177         LASSERT(qunit_cachep == NULL);
178         qunit_cachep = cfs_mem_cache_create("ll_qunit_cache",
179                                             sizeof(struct lustre_qunit),
180                                             0, 0);
181         if (!qunit_cachep)
182                 RETURN(-ENOMEM);
183
184         spin_lock(&qunit_hash_lock);
185         for (i = 0; i < NR_DQHASH; i++)
186                 CFS_INIT_LIST_HEAD(qunit_hash + i);
187         spin_unlock(&qunit_hash_lock);
188         RETURN(0);
189 }
190
191 static inline int
192 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
193              __attribute__((__const__));
194
195 static inline int
196 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
197 {
198         unsigned int id = qdata->qd_id;
199         unsigned int type = QDATA_IS_GRP(qdata);
200
201         unsigned long tmp = ((unsigned long)qctxt >> L1_CACHE_SHIFT) ^ id;
202         tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
203         return tmp;
204 }
205
206 /* caller must hold qunit_hash_lock */
207 static inline struct lustre_qunit *find_qunit(unsigned int hashent,
208                                               struct lustre_quota_ctxt *qctxt,
209                                               struct qunit_data *qdata)
210 {
211         struct lustre_qunit *qunit = NULL;
212         struct qunit_data *tmp;
213
214         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
215         list_for_each_entry(qunit, qunit_hash + hashent, lq_hash) {
216                 tmp = &qunit->lq_data;
217                 if (qunit->lq_ctxt == qctxt &&
218                     qdata->qd_id == tmp->qd_id &&
219                     (qdata->qd_flags & LQUOTA_QUNIT_FLAGS) ==
220                     (tmp->qd_flags & LQUOTA_QUNIT_FLAGS))
221                         return qunit;
222         }
223         return NULL;
224 }
225
226 /* check_cur_qunit - check the current usage of qunit.
227  * @qctxt: quota context
228  * @qdata: the type of quota unit to be checked
229  *
230  * return: 1 - need acquire qunit;
231  *         2 - need release qunit;
232  *         0 - need do nothing.
233  *       < 0 - error.
234  */
235 static int
236 check_cur_qunit(struct obd_device *obd,
237                 struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
238 {
239         struct super_block *sb = qctxt->lqc_sb;
240         unsigned long qunit_sz, tune_sz;
241         __u64 usage, limit, limit_org, pending_write = 0;
242         long long record = 0;
243         struct obd_quotactl *qctl;
244         struct lustre_qunit_size *lqs = NULL;
245         int ret = 0;
246         ENTRY;
247
248         if (!sb_any_quota_enabled(sb))
249                 RETURN(0);
250
251         spin_lock(&qctxt->lqc_lock);
252         if (!qctxt->lqc_valid){
253                 spin_unlock(&qctxt->lqc_lock);
254                 RETURN(0);
255         }
256         spin_unlock(&qctxt->lqc_lock);
257
258         OBD_ALLOC_PTR(qctl);
259         if (qctl == NULL)
260                 RETURN(-ENOMEM);
261
262         /* get fs quota usage & limit */
263         qctl->qc_cmd = Q_GETQUOTA;
264         qctl->qc_id = qdata->qd_id;
265         qctl->qc_type = QDATA_IS_GRP(qdata);
266         ret = fsfilt_quotactl(obd, sb, qctl);
267         if (ret) {
268                 if (ret == -ESRCH)      /* no limit */
269                         ret = 0;
270                 else
271                         CERROR("can't get fs quota usage! (rc:%d)\n", ret);
272                 GOTO(out, ret);
273         }
274
275         if (QDATA_IS_BLK(qdata)) {
276                 usage = qctl->qc_dqblk.dqb_curspace;
277                 limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
278         } else {
279                 usage = qctl->qc_dqblk.dqb_curinodes;
280                 limit = qctl->qc_dqblk.dqb_ihardlimit;
281         }
282
283         /* ignore the no quota limit case; and it can avoid creating
284          * unnecessary lqs for uid/gid */
285         if (!limit)
286                 GOTO(out, ret = 0);
287
288  search_lqs:
289         quota_search_lqs(qdata, NULL, qctxt, &lqs);
290         if (!lqs) {
291                 CDEBUG(D_QUOTA, "Can't find the lustre qunit size!\n");
292                 ret = quota_create_lqs(qdata, NULL, qctxt, &lqs);
293                 if (ret == -EALREADY) {
294                         ret = 0;
295                         goto search_lqs;
296                 }
297                 if (ret < 0)
298                         GOTO (out, ret);
299         }
300         spin_lock(&lqs->lqs_lock);
301
302         if (QDATA_IS_BLK(qdata)) {
303                 qunit_sz = lqs->lqs_bunit_sz;
304                 tune_sz  = lqs->lqs_btune_sz;
305                 pending_write = lqs->lqs_bwrite_pending;
306                 record   = lqs->lqs_blk_rec;
307                 LASSERT(!(qunit_sz % QUOTABLOCK_SIZE));
308         } else {
309                 /* we didn't need change inode qunit size now */
310                 qunit_sz = lqs->lqs_iunit_sz;
311                 tune_sz  = lqs->lqs_itune_sz;
312                 pending_write = lqs->lqs_iwrite_pending;
313                 record   = lqs->lqs_ino_rec;
314         }
315
316         /* we don't count the MIN_QLIMIT */
317         if ((limit == MIN_QLIMIT && !QDATA_IS_BLK(qdata)) ||
318             (toqb(limit) == MIN_QLIMIT && QDATA_IS_BLK(qdata)))
319                 limit = 0;
320
321         usage += pending_write;
322         limit_org = limit;
323         /* when a releasing quota req is sent, before it returned
324            limit is assigned a small value. limit will overflow */
325         if (limit + record < 0)
326                 usage -= record;
327         else
328                 limit += record;
329
330         LASSERT(qdata->qd_count == 0);
331         if (limit <= usage + tune_sz) {
332                 while (qdata->qd_count + limit <=
333                        usage + tune_sz)
334                         qdata->qd_count += qunit_sz;
335                 ret = 1;
336         } else if (limit > usage + qunit_sz + tune_sz &&
337                    limit_org > qdata->qd_count + qunit_sz) {
338                 while (limit - qdata->qd_count > usage + qunit_sz + tune_sz &&
339                        limit_org > qdata->qd_count + qunit_sz)
340                         qdata->qd_count += qunit_sz;
341                 ret = 2;
342                 /* if there are other pending writes for this uid/gid, releasing
343                  * quota is put off until the last pending write b=16645 */
344                 if (ret == 2 && pending_write) {
345                         CDEBUG(D_QUOTA, "delay quota release\n");
346                         ret = 0;
347                 }
348         }
349         CDEBUG(D_QUOTA, "type: %c, limit: "LPU64", usage: "LPU64
350                ", pending_write: "LPU64", record: "LPD64
351                ", qunit_sz: %lu, tune_sz: %lu, ret: %d.\n",
352                QDATA_IS_BLK(qdata) ? 'b' : 'i', limit, usage, pending_write,
353                record, qunit_sz, tune_sz, ret);
354         LASSERT(ret == 0 || qdata->qd_count);
355
356         spin_unlock(&lqs->lqs_lock);
357         lqs_putref(lqs);
358         EXIT;
359  out:
360         OBD_FREE_PTR(qctl);
361         return ret;
362 }
363
364 /**
365  * Compute the remaining quota for certain gid or uid b=11693
366  */
367 int compute_remquota(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
368                      struct qunit_data *qdata, int isblk)
369 {
370         struct super_block *sb = qctxt->lqc_sb;
371         __u64 usage, limit;
372         struct obd_quotactl *qctl;
373         int ret = QUOTA_RET_OK;
374         ENTRY;
375
376         if (!sb_any_quota_enabled(sb))
377                 RETURN(QUOTA_RET_NOQUOTA);
378
379         /* ignore root user */
380         if (qdata->qd_id == 0 && QDATA_IS_GRP(qdata) == USRQUOTA)
381                 RETURN(QUOTA_RET_NOLIMIT);
382
383         OBD_ALLOC_PTR(qctl);
384         if (qctl == NULL)
385                 RETURN(-ENOMEM);
386
387         /* get fs quota usage & limit */
388         qctl->qc_cmd = Q_GETQUOTA;
389         qctl->qc_id = qdata->qd_id;
390         qctl->qc_type = QDATA_IS_GRP(qdata);
391         ret = fsfilt_quotactl(obd, sb, qctl);
392         if (ret) {
393                 if (ret == -ESRCH)      /* no limit */
394                         ret = QUOTA_RET_NOLIMIT;
395                 else
396                         CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)",
397                                ret);
398                 GOTO(out, ret);
399         }
400
401         usage = isblk ? qctl->qc_dqblk.dqb_curspace :
402                 qctl->qc_dqblk.dqb_curinodes;
403         limit = isblk ? qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS :
404                 qctl->qc_dqblk.dqb_ihardlimit;
405         if (!limit){            /* no limit */
406                 ret = QUOTA_RET_NOLIMIT;
407                 GOTO(out, ret);
408         }
409
410         if (limit >= usage)
411                 qdata->qd_count = limit - usage;
412         else
413                 qdata->qd_count = 0;
414         EXIT;
415 out:
416         OBD_FREE_PTR(qctl);
417         return ret;
418 }
419
420 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
421                                         struct qunit_data *qdata, int opc)
422 {
423         struct lustre_qunit *qunit = NULL;
424         ENTRY;
425
426         OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
427         if (qunit == NULL)
428                 RETURN(NULL);
429
430         CFS_INIT_LIST_HEAD(&qunit->lq_hash);
431         init_waitqueue_head(&qunit->lq_waitq);
432         atomic_set(&qunit->lq_refcnt, 1);
433         qunit->lq_ctxt = qctxt;
434         memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
435         qunit->lq_opc = opc;
436         qunit->lq_lock = SPIN_LOCK_UNLOCKED;
437         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
438         RETURN(qunit);
439 }
440
441 static inline void free_qunit(struct lustre_qunit *qunit)
442 {
443         OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
444 }
445
446 static inline void qunit_get(struct lustre_qunit *qunit)
447 {
448         atomic_inc(&qunit->lq_refcnt);
449 }
450
451 static void qunit_put(struct lustre_qunit *qunit)
452 {
453         LASSERT(atomic_read(&qunit->lq_refcnt));
454         if (atomic_dec_and_test(&qunit->lq_refcnt))
455                 free_qunit(qunit);
456 }
457
458 /* caller must hold qunit_hash_lock and release ref of qunit after using it */
459 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
460                                             struct qunit_data *qdata)
461 {
462         unsigned int hashent = qunit_hashfn(qctxt, qdata);
463         struct lustre_qunit *qunit;
464         ENTRY;
465
466         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
467         qunit = find_qunit(hashent, qctxt, qdata);
468         if (qunit)
469                 qunit_get(qunit);
470         RETURN(qunit);
471 }
472
473 static void
474 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
475 {
476         struct list_head *head;
477
478         LASSERT(list_empty(&qunit->lq_hash));
479         qunit_get(qunit);
480         head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
481         list_add(&qunit->lq_hash, head);
482         QUNIT_SET_STATE(qunit, QUNIT_IN_HASH);
483 }
484
485 static void compute_lqs_after_removing_qunit(struct lustre_qunit *qunit)
486 {
487         struct lustre_qunit_size *lqs = NULL;
488
489         quota_search_lqs(&qunit->lq_data, NULL, qunit->lq_ctxt, &lqs);
490         if (lqs) {
491                 spin_lock(&lqs->lqs_lock);
492                 if (qunit->lq_opc == QUOTA_DQACQ)
493                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 1);
494                 if (qunit->lq_opc == QUOTA_DQREL)
495                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 0);
496                 spin_unlock(&lqs->lqs_lock);
497                 /* this is for quota_search_lqs */
498                 lqs_putref(lqs);
499                 /* this is for schedule_dqacq */
500                 lqs_putref(lqs);
501         }
502
503 }
504
505 static void remove_qunit_nolock(struct lustre_qunit *qunit)
506 {
507         LASSERT(!list_empty(&qunit->lq_hash));
508         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
509
510         list_del_init(&qunit->lq_hash);
511         QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
512         qunit_put(qunit);
513 }
514
515 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
516                                  (limit = count) : (limit += count)
517
518
519 static inline int is_master(struct lustre_quota_ctxt *qctxt)
520 {
521         return qctxt->lqc_handler ? 1 : 0;
522 }
523
524 static int
525 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
526                struct qunit_data *qdata, int opc, int wait,
527                struct obd_trans_info *oti);
528
529 static int
530 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
531                  struct qunit_data *qdata, int rc, int opc)
532 {
533         struct lustre_qunit *qunit = NULL;
534         struct super_block *sb = qctxt->lqc_sb;
535         int err = 0;
536         struct quota_adjust_qunit *oqaq = NULL;
537         int rc1 = 0;
538         ENTRY;
539
540         LASSERT(qdata);
541         QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
542                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
543
544         /* update local operational quota file */
545         if (rc == 0) {
546                 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
547                 struct obd_quotactl *qctl;
548                 __u64 *hardlimit;
549
550                 OBD_ALLOC_PTR(qctl);
551                 if (qctl == NULL)
552                         GOTO(out, err = -ENOMEM);
553
554                 /* acq/rel qunit for specified uid/gid is serialized,
555                  * so there is no race between get fs quota limit and
556                  * set fs quota limit */
557                 qctl->qc_cmd = Q_GETQUOTA;
558                 qctl->qc_id = qdata->qd_id;
559                 qctl->qc_type = QDATA_IS_GRP(qdata);
560                 err = fsfilt_quotactl(obd, sb, qctl);
561                 if (err) {
562                         CERROR("error get quota fs limit! (rc:%d)\n", err);
563                         GOTO(out_mem, err);
564                 }
565
566                 if (QDATA_IS_BLK(qdata)) {
567                         qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
568                         hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
569                 } else {
570                         qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
571                         hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
572                 }
573
574                 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
575
576                 if (*hardlimit == 0)
577                         goto out_mem;
578
579                 switch (opc) {
580                 case QUOTA_DQACQ:
581                         INC_QLIMIT(*hardlimit, count);
582                         break;
583                 case QUOTA_DQREL:
584                         LASSERTF(count < *hardlimit,
585                                  "id(%u) flag(%u) type(%c) isblk(%c) "
586                                  "count("LPU64") qd_qunit("LPU64") "
587                                  "hardlimit("LPU64").\n",
588                                  qdata->qd_id, qdata->qd_flags,
589                                  QDATA_IS_GRP(qdata) ? 'g' : 'u',
590                                  QDATA_IS_BLK(qdata) ? 'b': 'i',
591                                  qdata->qd_count, qdata->qd_qunit, *hardlimit);
592                         *hardlimit -= count;
593                         break;
594                 default:
595                         LBUG();
596                 }
597
598                 /* clear quota limit */
599                 if (count == 0)
600                         *hardlimit = 0;
601
602                 qctl->qc_cmd = Q_SETQUOTA;
603                 err = fsfilt_quotactl(obd, sb, qctl);
604                 if (err)
605                         CERROR("error set quota fs limit! (rc:%d)\n", err);
606
607                 QDATA_DEBUG(qdata, "%s completion\n",
608                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
609 out_mem:
610                 OBD_FREE_PTR(qctl);
611         } else if (rc == -EDQUOT) {
612                 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
613         } else if (rc == -EBUSY) {
614                 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
615         } else {
616                 CERROR("acquire qunit got error! (rc:%d)\n", rc);
617         }
618 out:
619         /* remove the qunit from hash */
620         spin_lock(&qunit_hash_lock);
621
622         qunit = dqacq_in_flight(qctxt, qdata);
623         /* this qunit has been removed by qctxt_cleanup() */
624         if (!qunit) {
625                 spin_unlock(&qunit_hash_lock);
626                 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
627                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
628                 RETURN(err);
629         }
630
631         LASSERT(opc == qunit->lq_opc);
632         /* remove this qunit from lq_hash so that new processes cannot be added
633          * to qunit->lq_waiters */
634         remove_qunit_nolock(qunit);
635         spin_unlock(&qunit_hash_lock);
636
637         compute_lqs_after_removing_qunit(qunit);
638
639         if (rc == 0)
640                 rc = QUOTA_REQ_RETURNED;
641         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
642         /* wake up all waiters */
643         wake_up_all(&qunit->lq_waitq);
644
645         /* this is for dqacq_in_flight() */
646         qunit_put(qunit);
647         /* this is for alloc_qunit() */
648         qunit_put(qunit);
649         if (rc < 0 && rc != -EDQUOT)
650                  RETURN(err);
651
652         /* don't reschedule in such cases:
653          *   - acq/rel failure and qunit isn't changed,
654          *     but not for quota recovery.
655          *   - local dqacq/dqrel.
656          *   - local disk io failure.
657          */
658          OBD_ALLOC_PTR(oqaq);
659          if (!oqaq)
660                  RETURN(-ENOMEM);
661          qdata_to_oqaq(qdata, oqaq);
662          /* adjust the qunit size in slaves */
663          rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
664          OBD_FREE_PTR(oqaq);
665          if (rc1 < 0) {
666                  CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
667                  RETURN(rc1);
668          }
669          if (err || (rc < 0 && rc != -EBUSY && rc1 == 0) || is_master(qctxt))
670                 RETURN(err);
671
672         /* reschedule another dqacq/dqrel if needed */
673         qdata->qd_count = 0;
674         qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
675         rc1 = check_cur_qunit(obd, qctxt, qdata);
676         if (rc1 > 0) {
677                 int opc;
678                 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
679                 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
680                 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
681         }
682         RETURN(err);
683 }
684
685 struct dqacq_async_args {
686         struct lustre_quota_ctxt *aa_ctxt;
687         struct lustre_qunit *aa_qunit;
688 };
689
690 static int dqacq_interpret(const struct lu_env *env,
691                            struct ptlrpc_request *req, void *data, int rc)
692 {
693         struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
694         struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
695         struct obd_device_target *obt = qctxt->lqc_obt;
696         struct lustre_qunit *qunit = aa->aa_qunit;
697         struct obd_device *obd = req->rq_import->imp_obd;
698         struct qunit_data *qdata = NULL;
699         int rc1 = 0;
700         ENTRY;
701
702         LASSERT(req);
703         LASSERT(req->rq_import);
704
705         /* there are several forms of qunit(historic causes), so we need to
706          * adjust qunit from slaves to the same form here */
707         OBD_ALLOC(qdata, sizeof(struct qunit_data));
708         if (!qdata)
709                 RETURN(-ENOMEM);
710
711         down_read(&obt->obt_rwsem);
712         /* if a quota req timeouts or is dropped, we should update quota
713          * statistics which will be handled in dqacq_completion. And in
714          * this situation we should get qdata from request instead of
715          * reply */
716         rc1 = quota_get_qdata(req, qdata,
717                               (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
718                               QUOTA_IMPORT);
719         if (rc1 < 0) {
720                 DEBUG_REQ(D_ERROR, req,
721                           "error unpacking qunit_data(rc: %d)\n", rc1);
722                 GOTO(exit, rc = rc1);
723         }
724
725         QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
726         QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
727
728         if (qdata->qd_id != qunit->lq_data.qd_id ||
729             OBD_FAIL_CHECK(OBD_FAIL_QUOTA_RET_QDATA)) {
730                 CDEBUG(D_ERROR, "the returned qd_id isn't expected!"
731                        "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
732                        qunit->lq_data.qd_id);
733                 qdata->qd_id = qunit->lq_data.qd_id;
734                 rc = -EPROTO;
735         }
736         if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
737                 CDEBUG(D_ERROR, "the returned grp/usr isn't expected!"
738                        "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
739                        qunit->lq_data.qd_flags);
740                 if (QDATA_IS_GRP(&qunit->lq_data))
741                         QDATA_SET_GRP(qdata);
742                 else
743                         QDATA_CLR_GRP(qdata);
744                 rc = -EPROTO;
745         }
746         if (qdata->qd_count > qunit->lq_data.qd_count) {
747                 CDEBUG(D_ERROR, "the returned qd_count isn't expected!"
748                        "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
749                        qunit->lq_data.qd_count);
750                 rc = -EPROTO;
751         }
752
753         rc = dqacq_completion(obd, qctxt, qdata, rc,
754                               lustre_msg_get_opc(req->rq_reqmsg));
755
756 exit:
757         up_read(&obt->obt_rwsem);
758         OBD_FREE(qdata, sizeof(struct qunit_data));
759
760         RETURN(rc);
761 }
762
763 /**
764  * check if quota master is online
765  */
766 int check_qm(struct lustre_quota_ctxt *qctxt)
767 {
768         int rc;
769         ENTRY;
770
771         spin_lock(&qctxt->lqc_lock);
772         /* quit waiting when mds is back or qctxt is cleaned up */
773         rc = qctxt->lqc_import || !qctxt->lqc_valid;
774         spin_unlock(&qctxt->lqc_lock);
775
776         RETURN(rc);
777 }
778
779 /* wake up all waiting threads when lqc_import is NULL */
780 void dqacq_interrupt(struct lustre_quota_ctxt *qctxt)
781 {
782         struct lustre_qunit *qunit, *tmp;
783         int i;
784         ENTRY;
785
786         spin_lock(&qunit_hash_lock);
787         for (i = 0; i < NR_DQHASH; i++) {
788                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
789                         if (qunit->lq_ctxt != qctxt)
790                                 continue;
791
792                         /* Wake up all waiters. Do not change lq_state.
793                          * The waiters will check lq_rc which is kept as 0
794                          * if no others change it, then the waiters will return
795                          * -EAGAIN to caller who can perform related quota
796                          * acq/rel if necessary. */
797                         wake_up_all(&qunit->lq_waitq);
798                 }
799         }
800         spin_unlock(&qunit_hash_lock);
801         EXIT;
802 }
803
804 static int got_qunit(struct lustre_qunit *qunit)
805 {
806         struct lustre_quota_ctxt *qctxt = qunit->lq_ctxt;
807         int rc = 0;
808         ENTRY;
809
810         spin_lock(&qunit->lq_lock);
811         switch (qunit->lq_state) {
812         case QUNIT_IN_HASH:
813         case QUNIT_RM_FROM_HASH:
814                 break;
815         case QUNIT_FINISHED:
816                 rc = 1;
817                 break;
818         default:
819                 CERROR("invalid qunit state %d\n", qunit->lq_state);
820         }
821         spin_unlock(&qunit->lq_lock);
822
823         if (!rc) {
824                 spin_lock(&qctxt->lqc_lock);
825                 rc = !qctxt->lqc_import || !qctxt->lqc_valid;
826                 spin_unlock(&qctxt->lqc_lock);
827         }
828
829         RETURN(rc);
830 }
831
832 static int
833 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
834                struct qunit_data *qdata, int opc, int wait,
835                struct obd_trans_info *oti)
836 {
837         struct lustre_qunit *qunit, *empty;
838         struct l_wait_info lwi = { 0 };
839         struct ptlrpc_request *req;
840         struct dqacq_async_args *aa;
841         struct obd_import *imp = NULL;
842         struct lustre_qunit_size *lqs = NULL;
843         struct timeval work_start;
844         struct timeval work_end;
845         long timediff;
846         int rc = 0;
847         ENTRY;
848
849         LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
850         do_gettimeofday(&work_start);
851         if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL)
852                 RETURN(-ENOMEM);
853
854         spin_lock(&qunit_hash_lock);
855         qunit = dqacq_in_flight(qctxt, qdata);
856         if (qunit) {
857                 spin_unlock(&qunit_hash_lock);
858                 qunit_put(empty);
859
860                 goto wait_completion;
861         }
862         qunit = empty;
863         qunit_get(qunit);
864         insert_qunit_nolock(qctxt, qunit);
865         spin_unlock(&qunit_hash_lock);
866
867         quota_search_lqs(qdata, NULL, qctxt, &lqs);
868         if (lqs) {
869                 spin_lock(&lqs->lqs_lock);
870                 quota_compute_lqs(qdata, lqs, 1, (opc == QUOTA_DQACQ) ? 1 : 0);
871                 /* when this qdata returned from mds, it will call lqs_putref */
872                 lqs_getref(lqs);
873                 spin_unlock(&lqs->lqs_lock);
874                 /* this is for quota_search_lqs */
875                 lqs_putref(lqs);
876         } else {
877                 CDEBUG(D_ERROR, "Can't find the lustre qunit size!\n");
878         }
879
880         QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
881                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
882         /* master is going to dqacq/dqrel from itself */
883         if (is_master(qctxt)) {
884                 int rc2;
885                 QDATA_DEBUG(qdata, "local %s.\n",
886                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
887                 QDATA_SET_CHANGE_QS(qdata);
888                 rc = qctxt->lqc_handler(obd, qdata, opc);
889                 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
890                 /* this is for qunit_get() */
891                 qunit_put(qunit);
892
893                 do_gettimeofday(&work_end);
894                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
895                 if (opc == QUOTA_DQACQ)
896                         lprocfs_counter_add(qctxt->lqc_stats,
897                                             wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
898                                             timediff);
899                 else
900                         lprocfs_counter_add(qctxt->lqc_stats,
901                                             wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
902                                             timediff);
903                 RETURN(rc ? rc : rc2);
904         }
905
906         spin_lock(&qctxt->lqc_lock);
907         if (!qctxt->lqc_import) {
908                 spin_unlock(&qctxt->lqc_lock);
909                 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
910
911                 spin_lock(&qunit_hash_lock);
912                 remove_qunit_nolock(qunit);
913                 spin_unlock(&qunit_hash_lock);
914
915                 compute_lqs_after_removing_qunit(qunit);
916
917                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
918                 wake_up_all(&qunit->lq_waitq);
919
920                 /* this is for qunit_get() */
921                 qunit_put(qunit);
922                 /* this for alloc_qunit() */
923                 qunit_put(qunit);
924                 spin_lock(&qctxt->lqc_lock);
925                 if (wait && !qctxt->lqc_import) {
926                         spin_unlock(&qctxt->lqc_lock);
927
928                         LASSERT(oti && oti->oti_thread &&
929                                 oti->oti_thread->t_watchdog);
930
931                         lc_watchdog_disable(oti->oti_thread->t_watchdog);
932                         CDEBUG(D_QUOTA, "sleep for quota master\n");
933                         l_wait_event(qctxt->lqc_wait_for_qmaster,
934                                      check_qm(qctxt), &lwi);
935                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
936                         lc_watchdog_touch(oti->oti_thread->t_watchdog);
937                 } else {
938                         spin_unlock(&qctxt->lqc_lock);
939                 }
940
941                 RETURN(-EAGAIN);
942         }
943         imp = class_import_get(qctxt->lqc_import);
944         spin_unlock(&qctxt->lqc_lock);
945
946         /* build dqacq/dqrel request */
947         LASSERT(imp);
948
949         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_QUOTA_DQACQ,
950                                         LUSTRE_MDS_VERSION, opc);
951         class_import_put(imp);
952         if (req == NULL) {
953                 CDEBUG(D_ERROR, "Can't alloc request\n");
954                 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
955                 /* this is for qunit_get() */
956                 qunit_put(qunit);
957                 RETURN(-ENOMEM);
958         }
959
960         ptlrpc_request_set_replen(req);
961         req->rq_no_resend = req->rq_no_delay = 1;
962         rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
963         if (rc < 0) {
964                 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
965                 ptlrpc_req_finished(req);
966                 dqacq_completion(obd, qctxt, qdata, -EPROTO, opc);
967                 /* this is for qunit_get() */
968                 qunit_put(qunit);
969                 RETURN(rc);
970         }
971
972         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
973         aa = ptlrpc_req_async_args(req);
974         aa->aa_ctxt = qctxt;
975         aa->aa_qunit = qunit;
976
977         req->rq_interpret_reply = dqacq_interpret;
978         ptlrpcd_add_req(req, PSCOPE_OTHER);
979
980         QDATA_DEBUG(qdata, "%s scheduled.\n",
981                     opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
982 wait_completion:
983         if (wait && qunit) {
984                 struct qunit_data *p = &qunit->lq_data;
985
986                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
987                 l_wait_event(qunit->lq_waitq, got_qunit(qunit), &lwi);
988                 /* rc = -EAGAIN, it means the quota master isn't ready yet
989                  * rc = QUOTA_REQ_RETURNED, it means a quota req is finished;
990                  * rc = -EDQUOT, it means out of quota
991                  * rc = -EBUSY, it means recovery is happening
992                  * other rc < 0, it means real errors, functions who call
993                  * schedule_dqacq should take care of this */
994                 spin_lock(&qunit->lq_lock);
995                 rc = qunit->lq_rc;
996                 spin_unlock(&qunit->lq_lock);
997                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting. (rc:%d)\n",
998                        qunit, rc);
999         }
1000
1001         qunit_put(qunit);
1002         do_gettimeofday(&work_end);
1003         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1004         if (opc == QUOTA_DQACQ)
1005                 lprocfs_counter_add(qctxt->lqc_stats,
1006                                     wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1007                                     timediff);
1008         else
1009                 lprocfs_counter_add(qctxt->lqc_stats,
1010                                     wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1011                                     timediff);
1012
1013         RETURN(rc);
1014 }
1015
1016 int
1017 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
1018                    uid_t uid, gid_t gid, __u32 isblk, int wait,
1019                    struct obd_trans_info *oti)
1020 {
1021         int rc = 0, i = USRQUOTA;
1022         __u32 id[MAXQUOTAS] = { uid, gid };
1023         struct qunit_data qdata[MAXQUOTAS];
1024         ENTRY;
1025
1026         CLASSERT(MAXQUOTAS < 4);
1027         if (!sb_any_quota_enabled(qctxt->lqc_sb))
1028                 RETURN(0);
1029
1030         for (i = 0; i < MAXQUOTAS; i++) {
1031                 qdata[i].qd_id = id[i];
1032                 qdata[i].qd_flags = i;
1033                 if (isblk)
1034                         QDATA_SET_BLK(&qdata[i]);
1035                 qdata[i].qd_count = 0;
1036
1037                 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
1038                 if (rc > 0) {
1039                         int opc;
1040                         /* need acquire or release */
1041                         opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1042                         rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1043                                             wait,oti);
1044                         if (rc < 0)
1045                                 RETURN(rc);
1046                 } else if (wait == 1) {
1047                         /* when wait equates 1, that means mds_quota_acquire
1048                          * or filter_quota_acquire is calling it. */
1049                         rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1050                         if (rc < 0)
1051                                 RETURN(rc);
1052                 }
1053         }
1054
1055         RETURN(rc);
1056 }
1057
1058 int
1059 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1060                          unsigned short type, int isblk)
1061 {
1062         struct lustre_qunit *qunit = NULL;
1063         struct qunit_data qdata;
1064         struct timeval work_start;
1065         struct timeval work_end;
1066         long timediff;
1067         struct l_wait_info lwi = { 0 };
1068         int rc = 0;
1069         ENTRY;
1070
1071         do_gettimeofday(&work_start);
1072         qdata.qd_id = id;
1073         qdata.qd_flags = type;
1074         if (isblk)
1075                 QDATA_SET_BLK(&qdata);
1076         qdata.qd_count = 0;
1077
1078         spin_lock(&qunit_hash_lock);
1079         qunit = dqacq_in_flight(qctxt, &qdata);
1080         spin_unlock(&qunit_hash_lock);
1081
1082         if (qunit) {
1083                 struct qunit_data *p = &qunit->lq_data;
1084
1085                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1086                 l_wait_event(qunit->lq_waitq, got_qunit(qunit), &lwi);
1087                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting. (rc:%d)\n",
1088                        qunit, qunit->lq_rc);
1089                 /* keep same as schedule_dqacq() b=17030 */
1090                 spin_lock(&qunit->lq_lock);
1091                 rc = qunit->lq_rc;
1092                 spin_unlock(&qunit->lq_lock);
1093                 /* this is for dqacq_in_flight() */
1094                 qunit_put(qunit);
1095                 do_gettimeofday(&work_end);
1096                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1097                 lprocfs_counter_add(qctxt->lqc_stats,
1098                                     isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1099                                             LQUOTA_WAIT_PENDING_INO_QUOTA,
1100                                     timediff);
1101         } else {
1102                 do_gettimeofday(&work_end);
1103                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1104                 lprocfs_counter_add(qctxt->lqc_stats,
1105                                     isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1106                                             LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1107                                     timediff);
1108         }
1109
1110         RETURN(rc);
1111 }
1112
1113 int
1114 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1115 {
1116         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1117         struct obd_device_target *obt = &obd->u.obt;
1118         struct super_block *sb = obt->obt_sb;
1119         int rc = 0;
1120         ENTRY;
1121
1122         LASSERT(qctxt);
1123
1124         rc = ptlrpcd_addref();
1125         if (rc)
1126                 RETURN(rc);
1127
1128         cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1129         spin_lock_init(&qctxt->lqc_lock);
1130         spin_lock(&qctxt->lqc_lock);
1131         qctxt->lqc_handler = handler;
1132         qctxt->lqc_sb = sb;
1133         qctxt->lqc_obt = obt;
1134         qctxt->lqc_import = NULL;
1135         qctxt->lqc_recovery = 0;
1136         qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1137         qctxt->lqc_valid = 1;
1138         qctxt->lqc_cqs_boundary_factor = 4;
1139         qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1140         qctxt->lqc_cqs_least_iunit = 2;
1141         qctxt->lqc_cqs_qs_factor = 2;
1142         qctxt->lqc_flags = 0;
1143         QUOTA_MASTER_UNREADY(qctxt);
1144         qctxt->lqc_bunit_sz = default_bunit_sz;
1145         qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1146         qctxt->lqc_iunit_sz = default_iunit_sz;
1147         qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1148         qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1149                                           * after the last shrinking */
1150         qctxt->lqc_sync_blk = 0;
1151         spin_unlock(&qctxt->lqc_lock);
1152
1153         qctxt->lqc_lqs_hash = lustre_hash_init("LQS_HASH", 7, 7,
1154                                                &lqs_hash_ops, 0);
1155         if (!qctxt->lqc_lqs_hash) {
1156                 CERROR("initialize hash lqs for %s error!\n", obd->obd_name);
1157                 RETURN(-ENOMEM);
1158         }
1159
1160 #ifdef LPROCFS
1161         rc = lquota_proc_setup(obd, is_master(qctxt));
1162         if (rc)
1163                 CERROR("initialize proc for %s error!\n", obd->obd_name);
1164 #endif
1165
1166         RETURN(rc);
1167 }
1168
1169 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1170 {
1171         struct lustre_qunit *qunit, *tmp;
1172         struct list_head tmp_list;
1173         struct obd_device_target *obt = qctxt->lqc_obt;
1174         int i;
1175         ENTRY;
1176
1177         CFS_INIT_LIST_HEAD(&tmp_list);
1178
1179         spin_lock(&qctxt->lqc_lock);
1180         qctxt->lqc_valid = 0;
1181         spin_unlock(&qctxt->lqc_lock);
1182
1183         spin_lock(&qunit_hash_lock);
1184         for (i = 0; i < NR_DQHASH; i++) {
1185                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
1186                         if (qunit->lq_ctxt != qctxt)
1187                                 continue;
1188                         remove_qunit_nolock(qunit);
1189                         list_add(&qunit->lq_hash, &tmp_list);
1190                 }
1191         }
1192         spin_unlock(&qunit_hash_lock);
1193
1194         list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1195                 list_del_init(&qunit->lq_hash);
1196                 compute_lqs_after_removing_qunit(qunit);
1197
1198                 /* wake up all waiters */
1199                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1200                 wake_up_all(&qunit->lq_waitq);
1201                 qunit_put(qunit);
1202         }
1203
1204         down_write(&obt->obt_rwsem);
1205         lustre_hash_exit(qctxt->lqc_lqs_hash);
1206         qctxt->lqc_lqs_hash = NULL;
1207         up_write(&obt->obt_rwsem);
1208
1209         /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1210          * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1211         while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1212                 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1213                 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE,
1214                                      cfs_time_seconds(1));
1215         }
1216
1217         ptlrpcd_decref();
1218
1219 #ifdef LPROCFS
1220         if (lquota_proc_cleanup(qctxt))
1221                 CERROR("cleanup proc error!\n");
1222 #endif
1223
1224         EXIT;
1225 }
1226
1227 struct qslave_recov_thread_data {
1228         struct obd_device *obd;
1229         struct lustre_quota_ctxt *qctxt;
1230         struct completion comp;
1231 };
1232
1233 /* FIXME only recovery block quota by now */
1234 static int qslave_recovery_main(void *arg)
1235 {
1236         struct qslave_recov_thread_data *data = arg;
1237         struct obd_device *obd = data->obd;
1238         struct lustre_quota_ctxt *qctxt = data->qctxt;
1239         unsigned int type;
1240         int rc = 0;
1241         ENTRY;
1242
1243         ptlrpc_daemonize("qslave_recovd");
1244
1245         complete(&data->comp);
1246
1247         if (qctxt->lqc_recovery)
1248                 RETURN(0);
1249         qctxt->lqc_recovery = 1;
1250
1251         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1252                 struct qunit_data qdata;
1253                 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1254                 struct list_head id_list;
1255                 struct dquot_id *dqid, *tmp;
1256                 int ret;
1257
1258                 LOCK_DQONOFF_MUTEX(dqopt);
1259                 if (!sb_has_quota_enabled(qctxt->lqc_sb, type)) {
1260                         UNLOCK_DQONOFF_MUTEX(dqopt);
1261                         break;
1262                 }
1263
1264                 LASSERT(dqopt->files[type] != NULL);
1265                 CFS_INIT_LIST_HEAD(&id_list);
1266 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1267                 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1268 #else
1269                 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1270 #endif
1271                 UNLOCK_DQONOFF_MUTEX(dqopt);
1272                 if (rc)
1273                         CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1274
1275                 list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1276                         list_del_init(&dqid->di_link);
1277                         /* skip slave recovery on itself */
1278                         if (is_master(qctxt))
1279                                 goto free;
1280                         if (rc && rc != -EBUSY)
1281                                 goto free;
1282
1283                         qdata.qd_id = dqid->di_id;
1284                         qdata.qd_flags = type;
1285                         QDATA_SET_BLK(&qdata);
1286                         qdata.qd_count = 0;
1287
1288                         ret = check_cur_qunit(obd, qctxt, &qdata);
1289                         if (ret > 0) {
1290                                 int opc;
1291                                 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1292                                 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1293                                                     0, NULL);
1294                                 if (rc == -EDQUOT)
1295                                         rc = 0;
1296                         } else {
1297                                 rc = 0;
1298                         }
1299
1300                         if (rc)
1301                                 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
1302                                        "qslave recovery failed! (id:%d type:%d "
1303                                        " rc:%d)\n", dqid->di_id, type, rc);
1304 free:
1305                         kfree(dqid);
1306                 }
1307         }
1308
1309         qctxt->lqc_recovery = 0;
1310         RETURN(rc);
1311 }
1312
1313 void
1314 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1315 {
1316         struct qslave_recov_thread_data data;
1317         int rc;
1318         ENTRY;
1319
1320         if (!sb_any_quota_enabled(qctxt->lqc_sb))
1321                 goto exit;
1322
1323         data.obd = obd;
1324         data.qctxt = qctxt;
1325         init_completion(&data.comp);
1326
1327         rc = kernel_thread(qslave_recovery_main, &data, CLONE_VM|CLONE_FILES);
1328         if (rc < 0) {
1329                 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1330                 goto exit;
1331         }
1332         wait_for_completion(&data.comp);
1333 exit:
1334         EXIT;
1335 }
1336
1337
1338 /**
1339  * lqs<->qctxt hash operations
1340  */
1341
1342 /**
1343  * string hashing using djb2 hash algorithm
1344  */
1345 static unsigned
1346 lqs_hash(lustre_hash_t *lh, void *key, unsigned mask)
1347 {
1348         struct quota_adjust_qunit *lqs_key;
1349         unsigned hash;
1350         ENTRY;
1351
1352         LASSERT(key);
1353         lqs_key = (struct quota_adjust_qunit *)key;
1354         hash = (QAQ_IS_GRP(lqs_key) ? 5381 : 5387) * lqs_key->qaq_id;
1355
1356         RETURN(hash & mask);
1357 }
1358
1359 static int
1360 lqs_compare(void *key, struct hlist_node *hnode)
1361 {
1362         struct quota_adjust_qunit *lqs_key;
1363         struct lustre_qunit_size *q;
1364         int rc;
1365         ENTRY;
1366
1367         LASSERT(key);
1368         lqs_key = (struct quota_adjust_qunit *)key;
1369         q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1370
1371         spin_lock(&q->lqs_lock);
1372         rc = ((lqs_key->qaq_id == q->lqs_id) &&
1373               (QAQ_IS_GRP(lqs_key) == LQS_IS_GRP(q)));
1374         spin_unlock(&q->lqs_lock);
1375
1376         RETURN(rc);
1377 }
1378
1379 static void *
1380 lqs_get(struct hlist_node *hnode)
1381 {
1382         struct lustre_qunit_size *q = 
1383             hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1384         ENTRY;
1385
1386         atomic_inc(&q->lqs_refcount);
1387         CDEBUG(D_QUOTA, "lqs=%p refcount %d\n",
1388                q, atomic_read(&q->lqs_refcount));
1389
1390         RETURN(q);
1391 }
1392
1393 static void *
1394 lqs_put(struct hlist_node *hnode)
1395 {
1396         struct lustre_qunit_size *q = 
1397             hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1398         ENTRY;
1399
1400         LASSERT(atomic_read(&q->lqs_refcount) > 0);
1401         atomic_dec(&q->lqs_refcount);
1402         CDEBUG(D_QUOTA, "lqs=%p refcount %d\n",
1403                q, atomic_read(&q->lqs_refcount));
1404
1405         RETURN(q);
1406 }
1407
1408 static void
1409 lqs_exit(struct hlist_node *hnode)
1410 {
1411         struct lustre_qunit_size *q;
1412         ENTRY;
1413
1414         q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1415         /* 
1416          * Nothing should be left. User of lqs put it and
1417          * lqs also was deleted from table by this time
1418          * so we should have 0 refs.
1419          */
1420         LASSERTF(atomic_read(&q->lqs_refcount) == 0, 
1421                  "Busy lqs %p with %d refs\n", q,
1422                  atomic_read(&q->lqs_refcount));
1423         OBD_FREE_PTR(q);
1424         EXIT;
1425 }
1426
1427 static lustre_hash_ops_t lqs_hash_ops = {
1428         .lh_hash    = lqs_hash,
1429         .lh_compare = lqs_compare,
1430         .lh_get     = lqs_get,
1431         .lh_put     = lqs_put,
1432         .lh_exit    = lqs_exit
1433 };
1434 #endif /* HAVE_QUOTA_SUPPORT */